Noradrenaline-trajectory phenotypes in septic shock: derivation and external validation in two independent cohorts

Figures

• 

  Fig. 1. Median NA trajectories by cluster. Solid lines represent the cluster‐specific median dose (µg/kg/min) from ICU admission; shaded bands represent the interquartile range. a Sheba cohort trajectories, b MIMIC-IV cohort trajectories

• 

  Fig. 2. Survival by NA‐trajectory cluster and ΔRMST versus lowest mortality cluster (landmark 96 h). a Kaplan–Meier survival to 90 days by cluster—Sheba cohort. b Kaplan–Meier survival to 90 days—MIMIC-IV cohort. c Difference in restricted mean survival time (ΔRMST) at 90 days relative to the first, lowest mortality cluster—Sheba cohort. d ΔRMST at 90 days—MIMIC-IV cohort

• 

  Fig. 3. Top discriminating features and top mortality-associated features. a Top discriminating features of NA trajectories in Sheba and MIMIC-IV. Bar plot showing the eight highest ranking features for separating cluster phenotypes in each cohort, based on combined random forest importance and ANOVA F-statistics. Exposure-related variables, particularly cumulative AUC windows (0–2 to 0–10 days) and mid-course exposure (AUC 24–48 h and 48–72 h), consistently ranked highest in both datasets, indicating strong cross-cohort agreement on the dimensions that define trajectory phenotypes. b Top mortality-associated NA-trajectory features (96-h landmark, 90-day mortality). Forest plot showing hazard ratios per 1-SD increase for the eight most prognostic features shared by Sheba and MIMIC-IV, using linear HR scale. Cumulative exposure features (e.g., AUC 0–10 days, 0–5 days, 0–4 days, 0–3 days) and early mid-course exposure (AUC 48–72 h, 72–96 h) show the strongest mortality associations in both cohorts, with broadly consistent effect sizes. Vertical dashed line denotes HR = 1.0 (no association)

• 

  Fig. 4. Multivariate landmark Cox models (24–144 h). Forest plots of hazard ratios (HRs) and 95% confidence intervals from multivariable landmark Cox proportional hazards models estimated at prespecified landmarks (24, 48, 72, 96, 120, and 144 h after ICU admission) in a the Sheba derivation cohort and b the external MIMIC-IV validation cohort. Points denote HR estimates and horizontal bars denote 95% confidence intervals; colors indicate landmark time. Continuous predictors were standardized to 1 SD to enable comparison of effect magnitudes; time_over_ variables are scaled per + 0.1 absolute increase. The vertical dashed line indicates HR = 1.0. Across both cohorts, age and SOFA-24 h were consistently associated with increased mortality risk, and norepinephrine-trajectory features reflecting delayed escalation (t_to_max_d), rebound (rebound_cnt), and time above dose thresholds (time_over_0.10, time_over_0.45) showed reproducible adverse associations. Decomposition of exposure into non-overlapping AUC windows highlighted a stronger and more consistent association for later-phase burden (48–96 h) compared with earlier windows

• 

  Fig. 5. Flow from 48-h predicted phenotype to final trajectory phenotype assignment. Alluvial diagrams show the relationship between phenotype predicted from truncated 48-h norepinephrine exposure data and final phenotype assignment based on the full trajectory framework in the derivation cohort (a, Sheba full derivation cohort with out-of-fold predictions) and the external validation cohort (b, MIMIC-IV). Left-sided strata represent the phenotype predicted at 48 h; right-sided strata represent the final assigned phenotype. Ribbon widths are proportional to patient counts. Dominant straight flows indicate stable early classification, whereas crossing flows indicate directional misclassification, most prominently among intermediate phenotypes